How Longstaff–Schwartz Least-Squares Monte Carlo Algorithms Price American Options: Continuation Regression, Exercise Policies, Basis Risk, Look-Ahead Bias and Lower-Bound Validation

Reader question: Monte Carlo simulation moves naturally forward through time, while an American option requires a decision at each exercise date based on the value of waiting. How can a simulation algorithm work backward through those forward paths and estimate when early exercise is rational?

The Longstaff–Schwartz least-squares Monte Carlo (LSMC) method solves this by turning the unknown continuation value into a regression problem. At each exercise date, simulated future discounted cash flows are regressed on functions of the current state. The fitted conditional expectation is compared with immediate exercise value, creating an approximate stopping policy.

This article owns one precise computational job: regression-based approximation of conditional continuation values for American/Bermudan optimal stopping on simulated paths. It does not own ordinary Monte Carlo path generation, Andersen–Broadie dual upper bounds, finite-difference free-boundary PDEs, or a particular stochastic asset model.

This is public mathematical and computational education. It is not financial advice, an options-trading recommendation or a guarantee that a fitted exercise policy is optimal.

1. Why American exercise breaks ordinary Monte Carlo

A European derivative pays only at maturity, so ordinary Monte Carlo can estimate:

V₀ = E^Q[D(0,T)H_T].

An American option can be exercised at many times. At time t the holder compares:

immediate exercise value H_t

with:

continuation value C_t = E^Q[D(t,τ)H_τ | F_t]

under the optimal future stopping rule.

The difficulty is circular: to decide whether to continue now, we need to know the value of future optimal decisions.

2. Dynamic programming provides the structure

On discrete Bermudan exercise dates t₀<t₁<…<t_M, the value satisfies:

V_m(X_m) = max(H_m(X_m), C_m(X_m)),

where:

C_m(x)=E[D(t_m,t_m+1)V_m+1(X_m+1) | X_m=x].

At maturity:

V_M=H_M.

So the algorithm can work backward if it can approximate each conditional expectation C_m.

3. Least squares approximates the continuation function

Choose basis functions:

φ₀(x), φ₁(x), …, φ_K(x).

Approximate:

C_m(x) ≈ Σ_k=0^K β_m,kφ_k(x).

For simulated paths i=1,…,N, define the regression target Y_m⁽ⁱ⁾ as the discounted cash flow eventually received if that path follows the decisions already determined at later exercise dates.

The least-squares coefficients solve:

β̂_m = arg min_β Σ [Y_m⁽ⁱ⁾ − φ(X_m⁽ⁱ⁾)^Tβ]².

The fitted continuation estimate is then:

Ĉ_m⁽ⁱ⁾ = φ(X_m⁽ⁱ⁾)^Tβ̂_m.

4. Exercise rule

For each path at exercise date t_m:

exercise if H_m(X_m) > Ĉ_m(X_m).

Otherwise continue.

If exercise occurs, all later cash flows on that path are discarded and the current exercise payoff becomes the path’s relevant cash flow.

The algorithm then moves one exercise date backward and repeats.

5. Why Longstaff and Schwartz often regress only in-the-money paths

For a standard option with non-negative exercise payoff, an out-of-the-money path has immediate exercise value zero. It is normally irrational to exercise there when continuation has non-negative value.

Regressing only in-the-money paths focuses the approximation near the region where the exercise-versus-wait decision is economically relevant.

But this is not a universal rule. Contracts with costs, multiple rights, callable features or unusual payoffs may require a different regression sample definition.

6. The original method is a projection

Least squares does not discover the exact continuation function unless it lies in the chosen function space.

It computes the best sample approximation within the span of the basis:

span{φ₀,…,φ_K}.

This creates two different errors:

• approximation error: the basis cannot represent the true continuation function;
• estimation error: finite simulated data estimate the projection coefficients imperfectly.

Adding more basis functions can reduce approximation error while increasing estimation variance and numerical instability.

7. Basis functions are a model choice

Common one-dimensional choices include:

• monomials 1,x,x²,…;
• Laguerre polynomials;
• Hermite polynomials;
• Chebyshev polynomials after scaling;
• problem-specific payoff/state transformations.

The original Longstaff–Schwartz examples famously used polynomial/Laguerre-type bases, but there is no theorem that one named family is always best.

Diagnostic: repeat the valuation with several well-scaled basis families and compare policy/value stability.

8. State selection matters more than polynomial fashion

If the payoff is Markov in S_t, current spot may be enough.

If the contract is path dependent, continuation may depend on additional state variables such as:

• running average;
• running maximum/minimum;
• remaining exercise rights;
• stochastic variance;
• interest-rate factors;
• multiple asset prices;
• accrued coupon or contractual state.

Leaving a required state variable out of the regression creates an information-loss error that no increase in path count can repair.

9. A minimal American put example

For an American put:

H(S)=max(K−S,0).

At an exercise date, regress discounted future realized cash flows on a basis such as:

1, S, S².

If the fitted continuation value is below K−S, exercise.

The estimated exercise boundary is the state region where the immediate-payoff curve crosses the fitted continuation curve.

This boundary is a powerful diagnostic: for a standard put under a simple model, it should be economically smooth rather than oscillating wildly from one nearby date to the next.

10. Training a stopping policy and valuing it are different jobs

The same simulated paths are often used to fit regressions and calculate an in-sample price.

That creates a statistical dependence: each path helps determine a regression rule that is then applied to itself.

A cleaner validation separates:

1. training paths: estimate β̂ and the exercise policy;
2. fresh valuation paths: apply the frozen policy without refitting.

The fresh-path estimate is especially important because any admissible stopping policy has expected value no greater than the true optimal American/Bermudan value. It therefore provides a genuine lower bound, subject to unbiased simulation of that policy.

11. Look-ahead bias

If a path helps fit the continuation regression and is then judged using that same fitted regression, the decision contains a small amount of information from its own future cash flow.

This is a form of look-ahead bias.

Recent leave-one-out LSM research shows that the bias can be material when the number of regressors is large relative to the number of paths, and can make an in-sample LSM estimate too high.

Falsifier: compare ordinary in-sample LSM with a fresh-path policy valuation or a leave-one-out construction. A persistent positive gap is evidence of regression reuse bias.

12. The estimator is not simply “always downward biased”

Two statements are often confused:

Policy value on independent fresh paths: lower bound, because a suboptimal stopping rule cannot beat the optimal rule in expectation.

In-sample LSM price: can contain look-ahead bias and need not behave like a clean lower bound.

This distinction is essential when auditing published LSMC results.

13. Regression conditioning

Raw powers such as 1,S,S²,S³ can become highly collinear when S occupies a narrow range or a large numerical scale.

The design matrix:

X^TX

can become ill-conditioned, making coefficients unstable.

Useful controls include:

• centre and scale state variables;
• orthogonal polynomial bases;
• QR decomposition rather than normal equations;
• SVD diagnostics;
• regularisation when justified;
• dropping redundant basis terms.

Falsifier: monitor condition numbers and coefficient sensitivity. A stable price built from wildly unstable coefficients deserves investigation rather than automatic acceptance.

14. More basis functions can make the policy worse

Suppose N is fixed and K rises.

Training residual error will usually fall because the regression has more flexibility.

But continuation estimates near the exercise boundary can become noisy, changing many stopping decisions.

This is ordinary statistical overfitting translated into an optimal-stopping problem.

Diagnostic: evaluate candidate basis sizes on fresh paths, not by training R² alone.

15. Regression R² is not the main objective

The algorithm does not need the best global prediction of future cash flow everywhere.

It needs sufficient accuracy near the decision boundary where:

H(x) ≈ C(x).

A regression can have a high R² while making poor decisions near that boundary, or a modest R² while still producing an excellent stopping policy.

Therefore decision diagnostics matter more than a single fit statistic.

16. Exercise-boundary diagnostics

Useful checks include:

• fraction exercised at each date;
• state distribution of exercised versus continued paths;
• distance H−Ĉ near the boundary;
• boundary smoothness through time;
• policy stability under new random seeds;
• comparison with tree/PDE boundaries in low-dimensional benchmarks.

A sudden saw-tooth boundary often signals regression noise, state omission or too few paths.

17. Exercise-date discretisation

An American option is continuously exercisable; an LSM simulation normally uses a finite set of candidate dates.

With too few dates, the algorithm prices a Bermudan approximation rather than the continuous-exercise contract.

Adding more dates should generally increase the feasible stopping set and move the value toward the continuous-exercise limit, but it also increases regression work and can amplify policy-estimation noise.

Falsifier: refine exercise dates while increasing path/regression quality enough to separate time-discretisation error from regression noise.

18. Multi-dimensional advantage—and curse

Longstaff and Schwartz highlighted cases where simulation remains practical even with many risk factors, including a high-dimensional term-structure example.

This is a major advantage over full tensor PDE grids.

But the curse of dimensionality reappears in the regression:

• many state variables;
• many interaction terms;
• many basis functions;
• more paths needed to estimate coefficients.

Simulation avoids the tensor grid; it does not abolish the statistical dimension problem.

19. Path-dependent contracts

LSMC is powerful when a low-dimensional sufficient state can summarize path history.

For an Asian-style early-exercise contract, for example, state might include:

(S_t, A_t)

where A_t is the running average.

If full path history matters in a way that cannot be compressed well, basis regression becomes difficult and alternative approximators may be more appropriate.

20. Multiple exercise rights

Swing options and some real/energy contracts allow several exercises.

The state must then include remaining rights and contractual timing constraints.

Dynamic programming becomes:

value = max(exercise now + value with one fewer right, continue with current rights).

LSMC can still approximate continuation values, but the state/action space grows substantially.

21. Inputs and outputs

Inputs can include:

• risk-neutral stochastic model;
• exercise dates;
• payoff/contract state;
• discount factors;
• training path count;
• validation path count;
• basis functions;
• state-variable definition;
• regression solver;
• regularisation/scaling rules;
• random-number policy;
• exercise-policy tie rule.

Outputs can include:

• in-sample fitted value;
• out-of-sample policy value/lower bound;
• Monte Carlo standard error;
• regression coefficients by date;
• condition numbers;
• exercise fractions;
• exercise-boundary diagnostics;
• basis-sensitivity table;
• path-count convergence;
• dual upper bound when paired with another method.

22. Evidence polarity

Evidence for confidence includes:

• fresh-path value agrees with tree/PDE benchmarks in low dimensions;
• policy value converges as training and validation paths increase;
• results are stable across sensible basis families;
• exercise boundaries are economically smooth;
• design matrices are well conditioned after scaling;
• fresh-path standard errors shrink approximately as 1/√N;
• basis expansion improves or stabilises out-of-sample value before overfitting appears;
• a dual upper bound is close to the policy lower bound.

Evidence against confidence includes:

• in-sample value materially exceeds fresh-path policy value;
• exercise policy changes drastically with random seed;
• higher-order bases increase out-of-sample error;
• exercise boundary oscillates;
• regression rank/conditioning is poor;
• important state variables are omitted;
• value does not stabilise under exercise-date refinement;
• fresh-path policy value lies far below an independent benchmark.

23. Counterexample: too many regressors, too few paths

Suppose 40 basis terms are fitted using only 200 in-the-money paths at one exercise date.

The regression can closely fit noise. The fitted continuation surface bends around individual path outcomes and the same paths receive favourable in-sample exercise decisions.

Falsifier: increase path count, reduce basis dimension, and compare with leave-one-out or independent policy valuation.

24. Counterexample: omitted stochastic volatility state

An American option is priced under a stochastic-volatility model, but continuation is regressed only on S.

Two paths with the same spot but very different variance states can have different continuation values.

Falsifier: add variance state and interactions. If out-of-sample policy value improves materially, the one-state regression was structurally incomplete.

25. Counterexample: high training R² but bad exercise policy

Future cash flows are highly variable, and the regression fits their broad level well. But near H≈C, errors of only a few cents reverse exercise decisions.

Falsifier: stratify paths by |H−Ĉ| and evaluate policy disagreement under richer/independent continuation estimates. Boundary error is the relevant loss.

26. Counterexample: policy evaluated on training paths

An LSM model reports a value above a high-quality tree benchmark.

This does not prove the tree is wrong. The training paths may be leaking future cash-flow information into their own fitted continuation estimates.

Falsifier: freeze the policy and apply it to an independent path set. If the value drops below the benchmark, the original excess was look-ahead bias.

27. Counterexample: too few exercise dates

A one-year American put is allowed to exercise only monthly in the simulation.

The resulting object is monthly Bermudan, not fully American.

Falsifier: increase exercise frequency and benchmark the value sequence. If it moves materially, time discretisation remains significant.

28. Alternatives

Finite-difference/lattice methods: often superior in low-dimensional Markov settings with tractable state grids.

Stochastic mesh: estimates continuation through weighted cross-path transitions rather than least-squares basis projection.

Dual martingale methods: produce upper bounds and are excellent validation partners.

Non-parametric regression: kernels/local methods can reduce basis misspecification but face dimension problems.

Neural networks/sparse polynomial approximators: can represent high-dimensional continuation surfaces but introduce their own training and validation risks.

29. Connection to Andersen–Broadie

Andersen–Broadie primal–dual algorithms own the complementary upper-bound problem.

An LSM policy evaluated on fresh paths gives a lower bound. A good dual martingale can give an upper bound. A narrow gap is far more informative than one unbracketed point estimate.

30. Connection to ordinary Monte Carlo and MLMC

Monte Carlo pricing algorithms own the underlying path-simulation estimator.

Multilevel Monte Carlo owns resolution-based complexity reduction. Research combines multilevel ideas with regression-based optimal stopping, but the regression-policy error remains a separate layer.

31. Connection to finite differences

Finite-difference option-pricing algorithms provide an independent low-dimensional benchmark for American exercise and free-boundary behaviour.

32. Weak links

• wrong risk-neutral path model;
• missing state variables;
• poorly scaled basis functions;
• ill-conditioned regression;
• too many regressors relative to in-the-money paths;
• look-ahead bias from same-path fitting/valuation;
• too few exercise dates;
• discount-factor/date errors;
• exercise tie rules changing many borderline paths;
• validation performed only in sample.

33. What would falsify confidence?

Confidence should be withdrawn if a frozen-policy value fails known low-dimensional benchmarks; if results are unstable across random seeds/basis sets; if look-ahead diagnostics show material overvaluation; if condition numbers are extreme; if state augmentation materially changes the policy; or if a dual upper bound remains far above the lower bound even after numerical refinement.

34. Verification and update triggers

Preserve model version, path generator, exercise calendar, state definition, basis functions, scaling, regression solver, training/validation seeds, coefficients, boundary diagnostics and lower/upper-bound comparisons.

Revalidate when:

• the payoff or exercise rights change;
• the stochastic model changes;
• new state factors are added;
• basis dimension changes;
• exercise frequency changes;
• regression libraries/solvers change;
• path count changes materially;
• look-ahead gap or duality gap widens.

35. Primary and high-quality references

• Francis A. Longstaff and Eduardo S. Schwartz, Valuing American Options by Simulation: A Simple Least-Squares Approach, Review of Financial Studies, 2001.
• Lars Stentoft, Convergence of the Least Squares Monte Carlo Approach to American Option Valuation, Management Science, 2004.
• L. C. G. Rogers, Monte Carlo Valuation of American Options, Mathematical Finance, 2002, for the dual upper-bound perspective.
• Chenru Liu and Jaehyuk Choi, Leave-One-Out Least Squares Monte Carlo Algorithm for Pricing Bermudan Options, Journal of Futures Markets, 2024.
• Daniel Z. Zanger, Convergence of a Least-Squares Monte Carlo Algorithm for American Option Pricing with Dependent Sample Data, Mathematical Finance.

Educational boundary: LSMC converts early exercise into a statistical learning problem. A regression that fits past simulated cash flows is not automatically a validated stopping policy; policy value, decision-boundary stability and independent upper/lower bounds are the decisive tests.